DeepLatent is a unified latent variable modeling framework for analyzing large multimodal and multilingual datasets. It relies on variational inference using deep neural networks for estimation.

The package currently supports:

• Generic latent factor models
• Topic models: The latent variables are a mixture of topics within documents.
• Ideal point models: The latent variables are interpreted as ideological dimensions.

• Multilingual and multimodal support

  • Learn topics / ideal points across multiple modalities (e.g., texts and images, texts and votes, etc.)
  • Learn the weight of each modality in determining the latent variables per observation

• Flexible metadata handling:

  • prevalence: covariates that influence the latent variables
  • content: covariates that influence the response variables conditional on the latent variables (e.g., topic-word distributions)
  • labels: outcomes for classification or regression tasks
  • prediction: additional predictors for the labels

• Flexible input/output representations:

  • Document embeddings (for texts, images, audio-visual data)
  • Word frequencies (BoW)
  • Raw images
  • Discrete choice data
  • Voting records

• Learns topics on the simplex
• Supports dirichlet or logistic_normal priors (optionally conditioned on covariates)

• Learns unconstrained latent variables (ℝ️ⁿ) for ideal point modeling
• Designed for political texts, images, audio and video recordings, surveys, and votes
• Uses a gaussian prior (optionally conditioned on covariates)

pip install deeplatent

git clone https://github.com/PinchOfData/DeepLatent.git  
cd deeplatent
pip install -e .

git clone https://github.com/PinchOfData/DeepLatent.git 
cd deeplatent
python setup_dev.py

Supports text, embeddings, votes, and survey questions:

import sys
sys.path.append('../src/')

from corpus import Corpus

modalities = {
    "text": {
        "column": "doc_clean",
        "views": {
            "bow": {
                "type": "bow",
                "vectorizer": CountVectorizer()
            }
        }
    },
    "image": {
        "column": "image_path",
        "views": {
            "embedding": {
                "type": "embedding",
                "embed_fn": my_image_embedder
            }
        }
    }
}

my_dataset = Corpus(df, modalities=modalities)

Optionally include metadata:

• prevalence, content, labels, prediction

from models import GTM

model = GTM(
    n_topics=20, 
    doc_topic_prior="logistic_normal",
    ae_type="wae"
)

from models import IdealPointNN

model = IdealPointNN(
    n_ideal_points=1, # one-dimensional ideal point model
    ae_type="vae"
)

• get_topic_words() – top words per topic
• get_covariate_words() – word shifts by content covariates
• get_top_docs() – representative documents
• get_topic_word_distribution() – topic-word matrix
• get_covariate_word_distribution() – word shift matrix
• plot_topic_word_distribution() – word clouds / bar plots
• visualize_docs() – document embeddings (UMAP, t-SNE, PCA)
• visualize_words() – word embeddings
• visualize_topics() – topic embeddings

• get_ideal_points() – ℝ️ⁿ latent space
• get_predictions() – supervised output
• get_modality_weights() – fusion weights (PoE or gating)

Check out the example notebooks to get started.

Download sample data to run some notebooks: Congressional Speeches CSV

• Deep Latent Variable Models for Unstructured Data , Germain Gauthier, Philine Widmer, Elliott Ash (2025)

• The Neural Ideal Point Model , Germain Gauthier, Hugo Subtil, Philine Widmer (2025)

This package is under active development 🚧 — feedback and contributions welcome!